Skip to content

PaddleOCR-VL: Boosting Document Parsing Efficiency and Performance with Coarse-to-Fine Visual Processing

Conference: CVPR 2026 arXiv: 2603.24326 Code: https://github.com/PaddlePaddle/PaddleOCR Area: Multimodal VLM / Document Understanding Keywords: Document parsing, coarse-to-fine visual processing, vision-language model, OCR, visual token compression

TL;DR

This paper proposes PaddleOCR-VL, a coarse-to-fine document parsing framework. The coarse stage employs a lightweight VRFM module to identify effective visual regions, while the fine stage applies a compact 0.9B VLM to process only those regions. With minimal visual tokens and parameters, the framework achieves state-of-the-art performance on OmniDocBench v1.5, substantially reducing latency and resource consumption.

Background & Motivation

  1. Background: Document parsing is a critical technology for constructing LLM training corpora and RAG systems. High-resolution inputs are essential for document parsing but cause quadratic growth in the number of visual tokens.
  2. Limitations of Prior Work: Pipeline-based methods (detection + recognition + structure reconstruction) are prone to error propagation; general-purpose VLMs hallucinate on handwritten or highly structured documents; specialized VLMs either have large parameter counts or suffer from coordinate drift.
  3. Key Challenge: High resolution is necessary for fine-grained recognition → visual token count explodes → high computational cost. Yet effective information in documents is highly non-uniform in distribution—PPT slides contain only ~39% effective regions, and newspapers approximately ~60%.
  4. Goal: Eliminate visual redundancy while preserving high-resolution accuracy, thereby improving efficiency.
  5. Key Insight: The sparsity of effective visual regions—large portions of background and decorative elements carry no useful information.
  6. Core Idea: A coarse stage rapidly identifies effective regions (localization + contextual relation prediction), and a fine stage processes only those regions.

Method

Overall Architecture

The framework consists of two stages: (1) Coarse stage—VRFM (Valid Region Focusing Module) rapidly localizes document elements (text, formulas, tables, etc.); (2) Fine stage—PaddleOCR-VL-0.9B performs detailed recognition on the extracted effective regions. The decoupled design allows independent optimization of each stage.

Key Designs

  1. VRFM (Valid Region Focusing Module):

    • Function: Rapidly locates semantically effective regions within a document.
    • Mechanism: A lightweight detector that simultaneously predicts region locations and inter-region contextual relationships (reading order). Compared to general-purpose object detectors, VRFM is optimized for document elements, achieving higher speed and accuracy.
    • Design Motivation: Standard VLMs that directly process full-page images waste substantial computation on background regions. Locating regions prior to recognition significantly reduces the input volume for the fine stage.

  2. PaddleOCR-VL-0.9B:

    • Function: Performs detailed text/formula/table recognition on effective visual regions.
    • Mechanism: A compact VLM with only 0.9B parameters, guided by VRFM outputs to process cropped effective regions rather than full-page images. The deficit in model scale is compensated through a high-quality data pipeline (30M+ samples).
    • Design Motivation: Small model + precise input > large model + coarse-grained input. The 0.9B scale enables edge deployment.

  3. Large-Scale High-Quality Data Pipeline:

    • Function: Provides sufficient training signal for the small model.
    • Mechanism: Over 30 million broadly distributed samples are collected from public sources and synthetic data, constituting one of the key factors behind model performance.
    • Design Motivation: Smaller models have lower data efficiency and require more high-quality data to compensate.

Loss & Training

VRFM: standard detection loss + reading order prediction loss. VLM: standard language modeling loss.

Key Experimental Results

Main Results

Metric PaddleOCR-VL GOT-OCR Qwen2.5-VL-7B InternVL3
Text Score SOTA 2nd 3rd 4th
Formula Score SOTA 2nd 3rd 4th
Table Score SOTA 2nd 3rd 4th
Reading Order SOTA 2nd 3rd 4th
Visual Tokens Fewest More Many Most
Parameters 0.9B Larger 7B Larger

Ablation Study

Configuration Overall Score Notes
Full PaddleOCR-VL SOTA VRFM + 0.9B VLM
w/o VRFM (direct full-page) Drops + slower Token count explodes
w/o data pipeline Significant drop Data is critical for small models

Key Findings

  • VRFM reduces the number of effective tokens by 40–60% while simultaneously improving performance.
  • A 0.9B model focused on effective regions can outperform a 7B general-purpose model.
  • The data pipeline is one of the key factors enabling the 0.9B model to achieve SOTA performance.

Highlights & Insights

  • The design philosophy of "small model + precise input > large model + coarse-grained input" has important implications for resource-constrained deployment.
  • The coarse-stage cost of VRFM is far lower than that of the VLM, yet it yields substantial savings in fine-stage computation—yielding an ROI of approximately 10–100×.
  • The decoupled design allows VRFM and the VLM to be independently optimized and upgraded.

Limitations & Future Work

  • Detection errors in VRFM propagate to the fine stage (missed regions cannot be recognized).
  • Robustness to extremely complex layouts (e.g., nested tables + formulas + images) requires further improvement.
  • Reading order prediction may fail under multi-column layouts.
  • vs. GOT-OCR: GOT-OCR processes entire pages end-to-end with high computational cost; PaddleOCR-VL reduces redundancy through the coarse-to-fine strategy.
  • vs. Qwen2.5-VL: General-purpose VLMs have large parameter counts (7B+) and are over-engineered for document scenarios; PaddleOCR-VL achieves superior results with only 0.9B parameters.
  • vs. Traditional Pipelines: Traditional pipelines lack global semantic understanding; the VLM fine stage in PaddleOCR-VL provides semantic comprehension capability.

Rating

  • Novelty: ⭐⭐⭐⭐ The coarse-to-fine document parsing framework and VRFM design demonstrate practical innovation.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation on OmniDocBench across four dimensions, achieving SOTA on all.
  • Writing Quality: ⭐⭐⭐⭐ Efficiency analysis is intuitive and comparisons are clearly presented.
  • Value: ⭐⭐⭐⭐⭐ The open-source framework offers significant value to the document AI community.