Baidu Unlimited-OCR: One-Shot Long-Horizon Parsing

Unlimited-OCR enables efficient long-document parsing

Baidu has released Unlimited-OCR, a model designed for "one-shot long-horizon parsing." This system allows for the transcription and parsing of extensive documents—such as multi-page PDFs—in a single pass, overcoming the memory limitations typically associated with processing long sequences in vision-language models.

Traditional OCR pipelines often require "janky code" to chop documents into individual pages and glue the text back together. Unlimited-OCR aims to eliminate this fragmented approach by handling long-horizon content natively.

Technical Architecture: Reference Sliding Window Attention (R-SWA)

Unlimited-OCR addresses the primary bottleneck of long-document OCR: the linear growth of the Key-Value (KV) cache. In standard transformer models, the memory required to remember every previously generated token grows linearly $O(N)$, which often leads to VRAM exhaustion when processing 100-page documents.

To solve this, the model employs Reference Sliding Window Attention (R-SWA), which splits the AI's focus into two distinct paths:

  • Global Reference: The model maintains full, uncompromised sight of the original document image, ensuring it never loses the visual context of the source material.
  • Local Generation: The model restricts its memory of its own generated text to a tight, moving window (e.g., the last 128 words). It safely "forgets" older generated text to keep memory usage constant regardless of document length.

Implementation and Inference

Unlimited-OCR is built upon the foundations of Deepseek-OCR and PaddleOCR. It is available on GitHub and Hugging Face, supporting two primary inference configurations for images:

Image Configuration Modes

  • Gundam Mode: Optimized for specific dimensions (base_size=1024, image_size=640, crop_mode=True).
  • Base Mode: Standard dimensions (base_size=1024, image_size=1024, crop_mode=False). Multi-page and PDF parsing exclusively use the Base mode.

Deployment Options

Users can run inference via the Hugging Face Transformers library or through SGLang for high-performance serving. The SGLang implementation supports an OpenAI-compatible API, allowing for streaming requests and batch inference via a provided infer.py script.

Local Requirements

Tested requirements include Python 3.12.3 and CUDA 12.9, with core dependencies such as torch==2.10.0, transformers==4.57.1, and pymupdf==1.27.2.2 for PDF-to-image conversion.

Community Insights and Perspectives

While the release has generated interest, technical discussions highlight several critical considerations for the future of OCR:

"My attempts at using AI to do OCR have always resulted in invented artifacts, which is not production feasible... words that are supposed to be in other languages being automatically translated to English, which ruins the effect."

Critics and users have raised questions regarding how this model compares to established industry standards like ABBYY FineReader or newer transformer-based benchmarks like Infinity Parser 2. There is also a recurring discussion on the necessity of re-engineering the OCR engine itself versus focusing on post-processing and data extraction.

Additionally, some users noted that the current state of Optical Music Recognition (OMR) remains a "greenfield for AI," as current tools struggle with the rich notation and symbolic understanding required for sheet music, contrasting with the relative maturity of text-based OCR.

Sources

Related

  • Project
    baidu/Unlimited-OCR

    A multimodal OCR model for one-shot long-horizon parsing of complex documents, PDFs, and multi-page images.

  • Dispatch
    Open-Source OCR Models Guide – Choosing, Running, and Extending Modern Vision-Language OCR

    Hugging Face announced a comprehensive guide to open-weight OCR models, detailing their capabilities, benchmark performance, cost efficiency, and how to run them locally or via hosted endpoints.

  • Dispatch
    Hugging Face Accelerating Document AI

    Hugging Face provides a comprehensive guide to using open-source multimodal models to automate document classification, parsing, and visual question answering for enterprise workflows.

  • Project
    PaddlePaddle/PaddleOCR

    PaddleOCR is an open‑source OCR and document‑AI toolkit that converts images/PDFs into structured, LLM‑ready JSON or Markdown. It offers multilingual scene‑text models (PP‑OCRv6), a lightweight vision‑language parser (PaddleOCR‑VL‑1.6) for tables, formulas, and ancient scripts, and a structure‑aware pipeline (PP‑StructureV3). Models run on CPU, GPU, XPU, NPU or via ONNX/TensorRT, and integrate with RAG/agent platforms like Dify and RAGFlow.

  • Dispatch
    Hugging Face releases vdr-2b-multi-v1 multilingual visual document retrieval model

    Hugging Face has introduced vdr-2b-multi-v1, a multilingual embedding model for visual document retrieval that enables searching complex documents without OCR by encoding page screenshots into dense vectors.