baidu / Unlimited-OCR

Overview

Unlimited OCR is an open-weight vision-language OCR model from Baidu. It uses Reference Sliding Window Attention (R-SWA) to keep decoding memory stable while it reads long documents. That design lets it transcribe dozens of pages in a single run, instead of doing page-by-page OCR loops. You get a single, continuous text result that you can search, chunk, or feed into downstream systems.

What you can build

• Searchable archives for scanned reports, books, and multi-page PDFs
• Document ingestion for RAG, with cleaner reading order across pages
• Contract and compliance pipelines that need one continuous transcription
• Table and formula transcription workflows for analytics and note-taking
• Multilingual digitization for mixed-language documents
• Batch OCR for folders of page images

Inputs

• One document file or multiple page files. The model accepts PDFs and common image formats, including PDF, JPG, JPEG, PNG, WebP, BMP, HEIC, TIFF, and AVIF.
• An instruction prompt that tells the model what to produce (for example, document parsing vs plain text). A clear prompt helps it keep structure.
• A maximum output length limit. Increase it for long documents to avoid truncated results.
• A DPI setting used when turning PDF pages into images before OCR. Higher DPI can improve small text, but it increases compute and output length.

Outputs

• Extracted text as a single response.
• The text can include document structure when the prompt asks for it. This often looks like Markdown-style formatting with headings, lists, and line breaks.
• For complex content, the output may include table markup and math expressions (commonly as HTML-like tables and LaTeX-style formulas) when the model recognizes them.
• For multi-page inputs, the output is meant to stay continuous across pages, which helps with cross-page tables and references.

Recommended settings

• For long documents, allow up to 32,768 tokens of output when you need full transcription.
• For a single page image, use the single-image high-detail mode (the project’s “gundam” configuration) that crops and uses a 640 image size.
• For multi-page documents and PDFs, use the multi-page mode (the project’s “base” configuration) with a 1024 image size.
• When rasterizing PDFs, 300 DPI is a strong default for small fonts and dense scans.
• Enable repetition control with an n-gram size of 35. Use a smaller repetition window for single pages (128) and a larger window for multi-page documents (1024).

Limitations

• The standard long-context setup targets a 32K maximum length. If the output needs more, it will truncate.
• Multi-page and PDF parsing uses the 1024 image size configuration. The single-page cropped mode is not meant for multi-page batches.
• PDF inputs typically require converting each page to an image first. Your DPI choice changes both accuracy and cost.
• Low-quality inputs can hurt results. Blurry scans, heavy compression, skewed pages, dark shadows, and tiny fonts can cause missing text or broken layout.
• Documents with inconsistent structure can confuse reading order. This includes complex slide decks, multi-column layouts, and mixed rotated elements.

Safety & compliance

• OCR often captures sensitive data. Only process documents you have rights to use.
• Get consent before extracting personal data from IDs, resumes, medical forms, or financial records.
• Store outputs securely. Treat the extracted text as sensitive if the source document was sensitive.