comfyui-unlimited-ocr

ComfyUI-Unlimited-OCR

ComfyUI custom nodes for Unlimited-OCR by Baidu - a vision-language model for high-accuracy optical character recognition (OCR) and text extraction.

Unlimited-OCR is a one-shot long-horizon parsing model capable of extracting text from images with high accuracy.

References

| Resource | Link | |----------|------| | HuggingFace Model | baidu/Unlimited-OCR | | GitHub Repository | baidu/Unlimited-OCR | | Paper | arXiv:2606.23050 |

Features

• Automatic Model Download: Downloads the Unlimited-OCR model from HuggingFace on first use
• High-Accuracy OCR: Extract text from images with high accuracy
• Text Localization: Locate and extract text regions within images
• Configurable Inference: Adjustable max_length, temperature, no_repeat_ngram_size, ngram_window, crop_mode, base_size, and image_size
• Model Caching: Models are cached to avoid redundant loading across workflows
• Debug Visualization: Inspect model loading state and test OCR results

Nodes

Loader & Config

| Node | Display Name | Description | |------|-------|-------------| | UnlimitedOCRLoader | Load Unlimited OCR Model | Loads the Unlimited-OCR model from HuggingFace with configurable device, dtype, and trust_remote_code settings | | UnlimitedOCRConfig | Configure Inference | Creates inference configuration with parameters for max_length, temperature, no_repeat_ngram_size, ngram_window, crop_mode, base_size, and image_size |

Inference

| Node | Display Name | Description | |------|-------|-------------| | UnlimitedOCRInference | OCR / Text Extraction | Performs OCR on an image with a custom prompt to control the type of OCR output |

Debug

| Node | Display Name | Description | |------|-------|-------------| | UnlimitedOCRDebug | Debug Model | Inspects model loading state and runs test OCR with visualized results |

Installation

1. Clone the repository into your ComfyUI custom nodes directory:

   cd ComfyUI/custom_nodes
    git clone https://github.com/PsychoLogicAu/ComfyUI-Unlimited-OCR.git
   cd ComfyUI-Unlimited-OCR
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Restart ComfyUI

4. Load the model in your workflow:

   • Add a Load Unlimited OCR Model node
   • Configure model path (default: baidu/Unlimited-OCR), dtype, and trust_remote_code
   • The model will be downloaded from HuggingFace on first use

Usage

Basic OCR Workflow

1. Add a Load Unlimited OCR Model node and configure:

   • model_path: baidu/Unlimited-OCR (or a local path)
   • dtype: bfloat16 (recommended), float16, or float32
   • trust_remote_code: Enable trust remote code (default: true)

2. (Optional) Add a Configure Inference node to customize:

   • max_length: Maximum output length (default: 32768)
   • temperature: Sampling temperature (default: 0.0)
   • no_repeat_ngram_size: No-repeat n-gram size (default: 35)
   • ngram_window: N-gram window size (default: 128)
   • crop_mode: Enable cropping (default: true)
   • base_size: Base image size (default: 1024)
   • image_size: Image size for processing (default: 640)

3. Add an OCR / Text Extraction node:

   • Connect the model output and your image
   • Set a custom prompt to control the type of OCR output (e.g., "document parsing.", "extract text.")
   • The node will extract text and return annotated images with bounding boxes

4. The node outputs:

   • extracted_text: Cleaned extracted text
   • annotated_image: Image with bounding boxes drawn
   • debug_text: Raw model output with bounding box annotations

Document Understanding Workflow

Use the OCR / Text Extraction node to extract and understand document content:

1. Connect model and image inputs
2. Set the prompt to "document parsing." for general document understanding
3. The node processes the image and extracts all text content
4. Adjust max_length if processing complex documents

Output Format

OCR Result

The extracted_text output from the OCR / Text Extraction node contains the cleaned extracted text.

The annotated_image output contains the image with bounding boxes drawn on detected text regions.

The debug_text output contains the raw model output with bounding box annotations.

Model Information

• Model: Baidu Unlimited-OCR
• HuggingFace: baidu/Unlimited-OCR
• Upstream Repo: baidu/Unlimited-OCR
• Paper: Unlimited OCR Works

Requirements

• Python 3.8+
• PyTorch 2.0+
• transformers
• Pillow (PIL)
• torchvision
• torchtyping
• sentencepiece

License

This project is provided as-is for use with ComfyUI. The Unlimited-OCR model is licensed under the MIT License. See the HuggingFace model card for licensing details.

Troubleshooting

• Model download fails: Check your internet connection and ensure you have access to the HuggingFace model
• CUDA out of memory: Reduce max_new_tokens
• Poor OCR quality: Try a higher resolution image, adjust temperature, or increase max_new_tokens

Contributing

Contributions welcome! Please feel free to submit issues and pull requests.

Image Attribution

Input image in examples: Kaldari, CC0, via Wikimedia Commons