GOT-OCR2.0

GOT-OCR2.0 is a next-generation optical charACTer reCognition system designed not just to detect text in images, but to deeply underStand and fAIthfully reconstruct diverse visual-language content. Developed by researchers at the university of Science and Technology of China, it reimAGInes OCR as a multimodal language assistant: instead of treating documents as static pixel grids, it interprets them as rich Information sources—extracting not only plain text but also mathematical notation, chemical structures, musical staves, GEOmetric diagrams, and tabular data, all while preserving semantic structure and fORMatting intent.

What it does

GOT-OCR2.0 performs end-to-end document intelligence. Given an image—whether a scanned textbook page, a medical report, a shipping label, or a scientific figure—it generates accurate, semantically faithful text ouTPUt. Unlike traditional OCR engines, it doesn’t stop at character bounding boxes; it reasons across modalities, distinguishing between prose, equations, and symbols, and outputs clean, editable results in Markdown or LaTeX. Its architecture supports interactive use: users can specify regions via coordinates or color Prompts, enabling precise extraction from cluttered or multi-column layouts.

Key capabilities

• Multimodal recognition: Processes text, math formulas, molecular diagrams, sheet music, and geometric sketches in a unified Framework.
• Long-context decoding: Handles up to 8K tokens, making it suitable for multi-page reports, lengthy contracts, or dense academic papers.
• Dynamic resolution handling: Adapts to ultra-high-resolution inputs without preprocessing loss, critical for archival scans or micrOSCopy images.
• Interactive region control: Allows selective, coordinate- or color-guided extraction—IDEal for forms, invoices, or annotated figures.
• multi-language & multi-format support: Natively handles Chinese, English, and other languages, with output options including plain text, markdown, and LaTeX.

Who it is for

Researchers, developers, and knowledge workers who need reliable, structured text extraction from heterogeneous documents—especially those working with technical, scientific, or multilingual materials. It’s particularly valuable for digitizing legacy academic resources, automating medical record processing, paRSIng engineering schematics, or building custom document analysis pipelines.

How to get started

The model is open-sourced on GitHub and can be run locally or deployed via Python APIs. Users need bASIc familiarity with PyTorch and Hugging Face Transformers. Pretrained weights and inference scripts are publicly available; minimal GPU memory (e.g., 12GB VRAM) is recommended for full-resolution Inference. A Lightweight web dEMO may be available through community deployments, though official hosted access is not provided.

Pricing and access caveat

GOT-OCR2.0 is fully Open Source and free to use under its license—no credits, subscriptions, or trials APPly. However, Local Deployment requires technical setup and compatible hardware. cloud-based inference services using GOT-OCR2.0 may vary in availability, pricing, and latency depending on third-party providers. As an academic research release, official commercial support, SLAs, or Enterprise integrations are not offered.