DeepSeek-OCR Portable

This is an easy-to-use offline OCR tool based on the DeepSeek-OCR 1280×1280 mode. It allows users to perform OCR without installing Python, Miniconda, CUDA Toolkit, or configuring environment variables. Just double-click and enjoy the powerful OCR experience powered by deep learning.

---

📋 Features

• Performs OCR on images, including complex structures like tables, formulas, figures, and references
• Output formats:
  • Markdown OCR result (<filename>.md)
  • Annotated image with bounding boxes (<filename>_with_boxes.jpg)
• "Green software" mode — All dependencies and models require no manual downloads; fully offline after initial setup

---

🖥️ System Requirements

• Windows 10 or Windows 11
• NVIDIA GPU (≥ 4GB VRAM)
• NVIDIA Driver ≥ 560.35

---

🚀 How to Use

1. Double-click init.bat (First run requires downloading models and dependencies - may take significant time)
2. A file selection window will appear — select the image you want to OCR
3. After processing, two files will be generated in the original image's directory:
   • Markdown OCR result: <original_filename>.md
   • Image with bounding boxes: <original_filename>_with_boxes.jpg

No need to install Python, Miniconda, or configure environment variables — all dependencies are automatically resolved!

---

📁 Project Structure

DeepSeek-OCR Portable/
├── env/                      # Portable Python environment
├── models/
│   └── DeepSeek-OCR/         # DeepSeek OCR model files
├── init.bat                  # One-click launch script (double-click to run)
├── run_ocr.bat               # Quick offline launch script (requires pre-downloaded models)
├── requirements.txt          # Python dependencies list
├── required_model_files.json # Model file list
├── check_model_files.py      # Model file existence checker
├── download_model_files.py   # Model download script
├── run_ocr.py                # OCR core logic
├── README.md                 # Documentation
├── README_zh.md              # Chinese documentation
└── LICENSE                   # MIT License

---

⚠️ Notes

• Initial download may be slow (~10GB) — please be patient
• If encountering "out of memory" errors:
  • Close other GPU-intensive applications
  • Modify IMAGE_SIZE in run_ocr.py to 1024 or 640
• Currently Windows-only (no macOS/Linux support)
• NVIDIA GPU required (CUDA 12.8 based) — AMD GPUs or CPU execution not supported

---

📝 Example Output

Suppose you select an image named document.jpg. After OCR processing, the following files will be generated in the original image's directory:

• document.md — OCR result in Markdown format
• document_with_boxes.jpg — Original image with detection boxes overlaid

---

🧠 Technical Details

• Uses HuggingFace's transformers library to load local models
• Uses torch.bfloat16 to reduce GPU memory usage
• Uses tkinter for file selection UI
• All dependencies are bundled in the env/ directory for true portability

---

📌 Developer Notes

You can customize the OCR behavior by modifying these parameters in run_ocr.py:

PROMPT = "<image>\n<|grounding|>Convert the document to markdown with full structure, including "
IMAGE_SIZE = 1280
BASE_SIZE = 1280
CROP_MODE = False
SAVE_RESULTS = True
TEST_COMPRESS = False

---

❤️ Acknowledgments

This project is built based on DeepSeek-OCR. Special thanks to the DeepSeek team for open-sourcing this high-quality OCR model.

---

Feel free to contribute, report issues, or improve this project!