Nougat
An academic document OCR model released by Meta AI, built on a Transformer encoder-decoder architecture. It takes full-page document images as input and outputs structured Markdown text directly. The model is optimized for scientific papers, with particular support for mathematical formula recognition in LaTeX format.
Timeline (last 90 days)
arXiv的PDF解析是学术论文跨库检索的主要难点，公式、表格、双栏排版会导致通用解析器出错。
Nougat (Neural Optical Understanding for Academic Documents) was released by Meta AI in 2023 and is based on a Transformer encoder-decoder architecture.
Nougat takes a full-page document image as input and directly outputs structured Markdown text, supports differentiation between inline and display formulas, and excels at recognizing multi-line formulas, matrices, and alignment environments.
The main limitations of Nougat and pix2tex are their limited robustness to low-quality scans (severe skewing, dense smudges), requiring image preprocessing steps such as deskewing, contrast enhancement, and denoising before input.
Nougat is a document-specific model from Meta that excels at academic paper and LaTeX formula recognition, with training data sourced from millions of arXiv papers.
All Facts (5)
arXiv的PDF解析是学术论文跨库检索的主要难点，公式、表格、双栏排版会导致通用解析器出错。
50%UnverifiedNougat (Neural Optical Understanding for Academic Documents) was released by Meta AI in 2023 and is based on a Transformer encoder-decoder architecture.
50%UnverifiedNougat takes a full-page document image as input and directly outputs structured Markdown text, supports differentiation between inline and display formulas, and excels at recognizing multi-line formulas, matrices, and alignment environments.
50%UnverifiedThe main limitations of Nougat and pix2tex are their limited robustness to low-quality scans (severe skewing, dense smudges), requiring image preprocessing steps such as deskewing, contrast enhancement, and denoising before input.
50%UnverifiedNougat is a document-specific model from Meta that excels at academic paper and LaTeX formula recognition, with training data sourced from millions of arXiv papers.
50%