Unlimited OCR Works (arxiv.org)

A recent technical report introduces Unlimited OCR, a model aimed at enhancing Optical Character Recognition (OCR) capabilities while addressing efficiency issues associated with traditional end-to-end systems, like DeepSeek OCR. By integrating a novel Reference Sliding Window Attention (R-SWA) mechanism, Unlimited OCR significantly reduces memory consumption during decoding, facilitating the transcription of extensive documents—up to 32K tokens—in a single forward pass. This advancement not only accelerates the OCR process but also mirrors human-like efficiency in handling lengthy text. The significance of Unlimited OCR for the AI/ML community lies in its innovative approach to memory management and computation. Traditional models suffer from escalating memory demands as output sequences lengthen, which can hinder performance and scalability. The introduction of R-SWA, which maintains a constant KV cache, presents a versatile solution not only for OCR but also for other natural language processing tasks such as automatic speech recognition (ASR) and translation. With the model's code and weights made publicly available, this work paves the way for future research and applications, fostering improvements in various fields reliant on efficient text processing.