Super-OCRs-Demo

A Gradio-based demo application for comparing state-of-the-art OCR models: DeepSeek-OCR, Dots.OCR, HunyuanOCR, and Nanonets-OCR2-3B. Users can upload images, select models, apply custom prompts, and generate recognized text or visual grounding results. Supports tasks like free OCR, markdown conversion, figure parsing, and object location.

Features

• Multi-Model Comparison: Switch between DeepSeek-OCR (with resolution and task options), Dots.OCR, HunyuanOCR, and Nanonets-OCR2-3B for flexible OCR workflows.
• Image Upload and Processing: Supports direct upload or clipboard paste; handles various image formats with PIL.
• Customizable Prompts: Tailor queries for text extraction, detection, or specific tasks (e.g., "Extract all text" or "Locate the red car").
• DeepSeek-Specific Tools: Resolution presets (Tiny to Gundam), task types (Free OCR, Markdown, Parse Figure, Locate Object), and bounding box visualization.
• Advanced Generation Controls: Adjust max new tokens (up to 8192), temperature, top-p, and top-k for fine-tuned outputs.
• Streaming Output: Real-time text generation for Dots.OCR and Nanonets-OCR2-3B; non-streaming for others.
• Visual Results: DeepSeek outputs annotated images with bounding boxes or grounding visuals.
• Custom Theme: SteelBlueTheme for a modern, gradient-based UI with enhanced readability.
• Examples and Queueing: Built-in example images; supports queued inferences for up to 30 concurrent users.

Prerequisites

• Python 3.10 or higher.
• CUDA-compatible GPU (recommended for bfloat16 models; falls back to CPU).
• Git for cloning submodules.
• Hugging Face account (optional, for model caching via huggingface_hub).

Installation

1. Clone the repository:

   git clone https://github.com/PRITHIVSAKTHIUR/Super-OCRs-Demo.git
   cd Super-OCRs-Demo
   

2. Install dependencies: Create a requirements.txt file with the following content, then run:

   pip install -r requirements.txt
   

   requirements.txt content:

   flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu12torch2.6cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
   git+https://github.com/huggingface/transformers@82a06db03535c49aa987719ed0746a76093b1ec4
   git+https://github.com/huggingface/accelerate.git
   git+https://github.com/huggingface/diffusers.git
   git+https://github.com/huggingface/peft.git
   huggingface_hub
   gradio==5.49.1
   qwen-vl-utils
   sentencepiece
   opencv-python
   torch==2.6.0
   torchvision
   supervision
   matplotlib
   easydict
   kernels
   einops
   spaces
   addict
   hf_xet
   numpy
   av
   

3. Start the application:

   python app.py
   

   The demo launches at http://localhost:7860 (or the provided URL if using Spaces).

Usage

1. Select Model: Choose from the radio buttons (default: DeepSeek-OCR).

   • DeepSeek: Adjust resolution (e.g., "Gundam (Recommended)") and task (e.g., "Convert to Markdown").
   • Others: Use the custom prompt textbox for queries like "Detect and extract all text with coordinates."

2. Upload Image: Drag-and-drop or paste an image (supports examples like receipts, figures, or documents).

3. Configure Settings:

   • For "Locate Object" in DeepSeek, enter reference text (e.g., "the title").
   • Tune advanced sliders for generation quality.

4. Run Inference: Click "Perform OCR" to process. Outputs stream to the textbox (with copy button); DeepSeek may show an annotated image.

5. View Results:

   • Text: Raw OCR output, markdown, or formatted coordinates.
   • Image: Bounding boxes in red for detected elements (DeepSeek only).

Example Workflow

• Upload a receipt image.
• Select Dots.OCR, prompt: "Extract items and prices."
• Adjust temperature to 0.1 for deterministic results.
• Output: Structured text list.

Supported Models

Model Name	Key Capabilities	Notes
DeepSeek-OCR-Latest-BF16.I64	Free OCR, Markdown, Figure Parsing, Object Location	Visual grounding with bounding boxes; resolution presets.
Dots.OCR-Latest-BF16	General text extraction; streaming	Qwen-based; custom prompts for flexibility.
HunyuanOCR	Detection and recognition with coordinates	Tencent model; handles Chinese/English well.
Nanonets-OCR2-3B	High-accuracy extraction; streaming	Qwen2.5-VL; suitable for complex layouts.

Troubleshooting

• Model Loading Errors: Ensure CUDA is installed for GPU; use torch.float32 fallback if bfloat16 fails.
• Out of Memory: Reduce resolution in DeepSeek or max_new_tokens; clear cache with torch.cuda.empty_cache().
• Import Issues: Install spaces only if deploying to Hugging Face Spaces; mock it locally.
• Generation Loops: Hunyuan may repeat; cleaned automatically via clean_repeated_substrings.
• UI Visibility: Model changes toggle DeepSeek-specific groups dynamically.
• Queue Full: Increase max_size in demo.queue() for high traffic.

Contributing

Contributions welcome! Open issues for bugs or features (e.g., more models, export to JSON). Fork, branch, and PR with tests. Repository: https://github.com/PRITHIVSAKTHIUR/Super-OCRs-Demo.git

License

Apache License 2.0. See LICENSE for details.

Built by Prithiv Sakthi. Report issues via the repository.