Unlimited OCR: One-Shot Long-Horizon Parsing (github.com)

A team of researchers has announced the release of Unlimited-OCR, an advanced model designed to enhance document parsing capabilities in optical character recognition (OCR) tasks. Leveraging advancements from its predecessor, Deepseek-OCR, Unlimited-OCR boasts one-shot long-horizon parsing, which allows it to handle longer text sequences more effectively. This development is significant for the AI/ML community as it promises improved efficiency and accuracy in processing both single images and multi-page documents, including PDFs, making it a valuable tool for a variety of applications in digital document management and data extraction. The model can be run through Hugging Face transformers on NVIDIA GPUs, supporting configurations for different image sizes and extraction modes. Notably, it allows users to set parameters such as maximum length for output texts, n-gram sizes for reducing repetition, and specific configurations for single or multi-page parsing. Unlimited-OCR also integrates a custom logit processor that enhances its parsing capabilities, which could lead to improved user experience in applications requiring fast and reliable document conversion. The model is now available on ModelScope and its paper can be accessed on arXiv, underlining its potential impact on OCR methodologies and applications.