Skip to main content

Unlimited-OCR-2

Overview​

Unlimited-OCR-2 parses document images into layout blocks with categories, coordinates, and recognized content. Its Sparse Reference Sliding Window Attention (SR-SWA) supports long-document decoding by maintaining a bounded set of working references. Multi-page inference uses ordered pages in one document request, retaining cross-page context and continuation relationships for text and tables.

Installation​

Follow the X-AnyLabeling-Server installation guide (English | 简体中文). Install the unlimited_ocr_2 extra for native Transformers inference, or the unlimited_ocr_2_sglang extra in the SGLang serving environment. These backends require different Transformers versions.

Enable unlimited_ocr_2 in the Server's configs/models.yaml. Configure the backend in configs/auto_labeling/unlimited_ocr_2.yaml:

  • backend: transformers: loads the pinned Hugging Face weights locally; model_path also accepts a local snapshot directory.
  • backend: sglang: connects to the customized upstream SGLang service through api_url. Set served_model to its served model name. An optional key can be supplied through SGLANG_API_KEY or api_key.

For SGLang deployment and page-window settings, see the Server deployment guide.

Usage​

  1. Configure the X-AnyLabeling remote server address to point to X-AnyLabeling-Server.
  2. Open the Document Parser (Ctrl+4) and select Unlimited-OCR-2 in Parsing model. This model uses the document panel, not the main-window annotation workflow.
  3. Click + New Parsing, select an image or PDF, and start parsing. The client renders PDF previews locally, then sends all pages in order in one document request. It does not run independent inference on each page.
  4. Review the source page and structured content side by side. Use the page controls to navigate PDFs and select blocks to inspect their source regions.
  5. Correct individual blocks, copy recognized content, or export the document/JSON using the panel's existing controls. The JSON preserves page geometry and global group IDs for cross-page continuation blocks.

PDF document parsing

Second-page table results

Notes​

  • Both Transformers and SGLang support joint multi-page requests. SGLang also supports long-document page windows with retained context between windows; the adapter consumes its streaming response and checks completion and page count before saving results.
  • max_new_tokens is the output budget per input page; the adapter multiplies it by the document's page count. A truncated response is reported as an error. crop_mode: null uses tiling for one page and base mode for multi-page input; true or false explicitly selects either mode.
  • The native backend is bounded by max_model_len and available GPU memory. Long documents should use the SGLang page-window scheduler. Window size affects memory use and how much cross-page context can be retained.
  • The panel's remote document timeout defaults to 3600 seconds and can be configured with remote_server_settings.document_timeout. Cancelling stops client-side waiting; inference already running on the server may continue.
  • Recognition quality depends on the input. For example, the upstream native model reads the handwritten E in the formula demo as S; the integration does not silently rewrite model predictions.