Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PaddleOCR-VL-1.5: Towards a Multi-Task 0.9B VLM for Robust In-the-Wild Document Parsing

Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, Zelun Zhang, Jiaxuan Liu, Xueqing Wang, Changda Zhou, Hongen Liu, Manhui Lin, Yue Zhang, Yubo Zhang, Yi Liu, Dianhai Yu, Yanjun Ma

TL;DR

PaddleOCR-VL-1.5 introduces a $0.9B$ ultra-compact Vision-Language Model with PP-DocLayoutV3 for unified, robust layout analysis and end-to-end reading order, plus an enhanced element recognition and text spotting capability. It achieves SOTA performance on OmniDocBench v1.5 (94.5%) and Real5-OmniDocBench (92.05%), and extends to Seal Recognition and Text Spotting with a 4-point text localization representation and a progressive, distortion-aware training regime. A new in-house Real5-OmniDocBench benchmark and Uncertainty-Aware Cluster Sampling (UACS) data curation strategy drive improved robustness to real-world distortions such as skew, warping, illumination, and screen photography, while maintaining strong efficiency (1.4335 pages/s on A100 with FastDeploy). The work demonstrates that a compact VLM can outperform much larger general-purpose models on document-centric tasks, enabling reliable deployment for document parsing, RAG systems, and downstream LLM-based workflows in unconstrained settings.

Abstract

We introduce PaddleOCR-VL-1.5, an upgraded model achieving a new state-of-the-art (SOTA) accuracy of 94.5% on OmniDocBench v1.5. To rigorously evaluate robustness against real-world physical distortions, including scanning, skew, warping, screen-photography, and illumination, we propose the Real5-OmniDocBench benchmark. Experimental results demonstrate that this enhanced model attains SOTA performance on the newly curated benchmark. Furthermore, we extend the model's capabilities by incorporating seal recognition and text spotting tasks, while remaining a 0.9B ultra-compact VLM with high efficiency. Code: https://github.com/PaddlePaddle/PaddleOCR

PaddleOCR-VL-1.5: Towards a Multi-Task 0.9B VLM for Robust In-the-Wild Document Parsing

TL;DR

PaddleOCR-VL-1.5 introduces a ultra-compact Vision-Language Model with PP-DocLayoutV3 for unified, robust layout analysis and end-to-end reading order, plus an enhanced element recognition and text spotting capability. It achieves SOTA performance on OmniDocBench v1.5 (94.5%) and Real5-OmniDocBench (92.05%), and extends to Seal Recognition and Text Spotting with a 4-point text localization representation and a progressive, distortion-aware training regime. A new in-house Real5-OmniDocBench benchmark and Uncertainty-Aware Cluster Sampling (UACS) data curation strategy drive improved robustness to real-world distortions such as skew, warping, illumination, and screen photography, while maintaining strong efficiency (1.4335 pages/s on A100 with FastDeploy). The work demonstrates that a compact VLM can outperform much larger general-purpose models on document-centric tasks, enabling reliable deployment for document parsing, RAG systems, and downstream LLM-based workflows in unconstrained settings.

Abstract

We introduce PaddleOCR-VL-1.5, an upgraded model achieving a new state-of-the-art (SOTA) accuracy of 94.5% on OmniDocBench v1.5. To rigorously evaluate robustness against real-world physical distortions, including scanning, skew, warping, screen-photography, and illumination, we propose the Real5-OmniDocBench benchmark. Experimental results demonstrate that this enhanced model attains SOTA performance on the newly curated benchmark. Furthermore, we extend the model's capabilities by incorporating seal recognition and text spotting tasks, while remaining a 0.9B ultra-compact VLM with high efficiency. Code: https://github.com/PaddlePaddle/PaddleOCR
Paper Structure (42 sections, 3 equations, 25 figures, 14 tables)

This paper contains 42 sections, 3 equations, 25 figures, 14 tables.

Table of Contents

  1. Introduction
  2. PaddleOCR-VL-1.5
  3. Architecture
  4. PP-DocLayoutV3: Unified Layout Analysis
  5. PaddleOCR-VL-1.5-0.9B: Element-level Recognition and Text Spotting
  6. Training Recipe
  7. Layout Analysis
  8. Element-level Recognition and Text Spotting
  9. Dataset
  10. Layout Analysis
  11. PaddleOCR-VL-1.5-0.9B
  12. Data Selection Strategy: Uncertainty-Aware Cluster Sampling
  13. Data Construction for New Capabilities
  14. Evaluation
  15. Document Parsing
  16. ...and 27 more sections

Figures (25)

  • Figure 1: Performance of PaddleOCR-VL-1.5 on OmniDocBench v1.5 and Real5-OmniDocBench.
  • Figure 2: The overview of PaddleOCR-VL-1.5.
  • Figure 3: The unified architecture of PP-DocLayoutV3, featuring parallel heads for instance segmentation and relational reading order prediction.
  • Figure 4: Samples of Real5-OmniDocBench.
  • Figure A1: The Markdown Output for Illumination.
  • ...and 20 more figures