### 📊 To Test ### ⭐ Exceptional - https://huggingface.co/lightonai/LightOnOCR-2-1B (20 s/page, excellent markdown and carot symbol for superscript for footnotes) - https://huggingface.co/PaddlePaddle/PaddleOCR-VL-1.5 (12.5 s/page, excellent text but no footnote superscript though, less vram) - https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B (8 s/page, excellent text but no footnote superscript, low vram, completely removes repetitive underscores though) ### ✅ Good - https://huggingface.co/zai-org/GLM-OCR (78 s/page, excellent text, no superscript for footnotes, shortens prolix underscores) - https://huggingface.co/nanonets/Nanonets-OCR2-3B (23 s/page, excellent text, superscript for footnotes, careful with max pixels) - https://huggingface.co/nanonets/Nanonets-OCR2-1.5B-exp (11.6 s/page, excellent markdown, superscript for footnotes, <footer> and <page_number> tags...not sure if beneficial ) - https://huggingface.co/florence-community/Florence-2-large (7.3 s/page, lower quality than tesseract, low vram, maybe useful to process limited bounding box images...) ### ❌ Unacceptable - https://huggingface.co/deepseek-ai/DeepSeek-OCR-2 (not compatible with Transformers 5+ yet) - https://huggingface.co/datalab-to/chandra (too large, didn't want to test) - https://huggingface.co/stepfun-ai/GOT-OCR2_0 (older, tested before, decided to not test again)
