README.md

metadata

title: Searchable PDF Generator with OCR (Hunyuan-DiT)
emoji: 📄
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.36.1
app_file: app.py
pinned: false

Searchable PDF Generator with OCR + Layout Preservation (Hunyuan-DiT)

This Hugging Face Space converts image-based PDFs and images into layout-preserving, searchable PDFs using the Hunyuan-DiT model for high-quality visual understanding and text extraction.

Objective

To develop a robust system that accepts PDFs or images, processes each page using Hunyuan-DiT for advanced OCR, extracts text with precise positional data, and reconstructs a searchable PDF that exactly mirrors the original visual layout, including page size, coordinates, fonts, bounding boxes, multi-column structures, tables, and embedded graphical elements.

Features

• Input Flexibility: Accepts PDF documents and various image formats (PNG, JPG, TIFF).
• High-Quality OCR: Utilizes Hunyuan-DiT for superior visual understanding and text extraction.
• Layout Preservation: Ensures the output PDF maintains the original document's layout, including:
  • Page size
  • Layout and coordinates
  • Approximate fonts
  • Bounding boxes for text
  • Multi-column structures
  • Tables and embedded graphical elements
• Searchable Output: Generates a multi-page searchable PDF with invisible text layers placed at exact coordinates over the original scanned image.
• Preview Functionality: Provides a preview of extracted OCR, highlighting bounding boxes.
• Downloadable Results: Offers a direct download link for the final searchable PDF.

Architecture

1. User Uploads: Users upload a PDF or image files via a Gradio interface.
2. PDF to Image Conversion: PDFs are converted into individual page images at 300–400 DPI.
3. Hunyuan-DiT OCR: Each page image is fed into the Hunyuan-DiT model for OCR and layout extraction, producing structured JSON output including words, bounding boxes, line grouping, reading order, block segmentation, and confidence scores.
4. PDF Assembly: An invisible text layer is meticulously assembled on top of the original image for each page.
5. Final PDF Generation: A multi-page searchable PDF is generated.
6. Download: The user can download the final searchable PDF.

Error Handling

The system includes robust error handling for:

• Corrupted PDF inputs.
• Multi-language input (handled by Hunyuan-DiT).
• Pages exceeding maximum size (>10,000 pixels in any dimension).

Setup and Usage

Prerequisites

Ensure you have Python 3.8+ installed.

Installation

1. Clone the repository:

   git clone <repository_url>
   cd <repository_name>
   

2. Create a virtual environment (recommended):

   python -m venv venv
   source venv/bin/activate  # On Windows: `venv\Scripts\activate`
   

3. Install the required dependencies:

   pip install -r requirements.txt
   

Running the Space

To launch the Gradio application locally:

python app.py

Then, open your web browser and navigate to the address provided by Gradio (usually http://127.0.0.1:7860).

Folder Structure

/
├── app.py
├── requirements.txt
├── utils/
│   ├── pdf_to_images.py
│   ├── hunyuan_ocr.py
│   └── pdf_builder.py
└── README.md