OCR-VLs WebUI

OCR Engines

• DeepSeekOCR: AI-powered OCR with advanced document understanding and markdown conversion
• PaddleOCR-VL: Document parsing model that converts documents to markdown format (install with: pip install 'paddleocr[doc-parser]')
• Gemini Flash 2.5: Google Gemini model for fast, high-quality Markdown conversion (set GEMINI_API_1..5 in .env)
• olmOCR: Vision-language model for document OCR (requires Python >=3.11)
• dots.ocr: Multilingual document parser with SOTA performance on layout detection and content recognition (install with: pip install qwen-vl-utils)

DeepSeekOCR Modes

• Gundam: 1024 base + 640 tiles with cropping - Best balance
• Tiny: 512×512, no crop - Fastest
• Small: 640×640, no crop - Quick
• Base: 1024×1024, no crop - Standard
• Large: 1280×1280, no crop - Highest quality

DeepSeekOCR Tasks

• Markdown: Convert document to structured markdown (grounding)
• Tables: Extract tables only as Markdown (grounding)
• Locate: Find specific text in image (grounding)
• Describe: General image description
• Custom: Your own prompt (add <|grounding|> for boxes)

PaddleOCR-VL

• Document parsing model that automatically converts documents to markdown
• Supports both images and PDFs

olmOCR

• Vision-language model based on Qwen2.5-VL-7B-Instruct
• Automatically converts documents to markdown format
• Supports both images and PDFs
• Model: allenai/olmOCR-2-7B-1025
• Requires Python >=3.11 - For Hugging Face Spaces, create a runtime.txt file with python-3.11 or higher

dots.ocr

• Multilingual document parser based on 1.7B LLM with SOTA performance
• Achieves state-of-the-art results for text, tables, and reading order
• Supports both images and PDFs
• Model: rednote-hilab/dots.ocr
• Requires transformers >= 4.47.0 - Please upgrade transformers if you see import errors

Gemini Flash 2.5

• Google Gemini model for fast, high-quality Markdown conversion