GOT-OCR2.0: Unified End-to-End OCR Model Released

Existing OCR approaches (OCR-1.0) have a pipeline structure that combines multiple modules such as text detection, recognition, and layout analysis, which has the limitation of accumulating errors between modules.

GOT-OCR2.0 proposes an End-to-End model that solves this by defining and processing all optical signals as a single 'character.' A single model with 5.8 hundred million (580M) parameters extracts text directly from images.

Key Features and Functions:

• Support for Various Formats: In addition to plain text, it uniformly recognizes formulas (LaTeX), molecular formulas (SMILES), tables (Markdown), sheet music, charts, and geometric figures.
• Structural Characteristics: It adopts a High-compression Encoder and Long-context Decoder structure to efficiently process document images with a large amount of information.
• Flexible Output: It can generate either simple text or structured data with formatting simply by adjusting the prompt.
• Fine-grained Control: It supports interactive OCR functionality that recognizes only a specific region of an image by specifying particular coordinates (Box) or colors (Color).

This model was developed by researchers at the University of Chinese Academy of Sciences (UCAS), and both the weights and training code are publicly available. It has also been merged into the Hugging Face Transformers library since February 2025, enabling batch inference.