How lukas-blecher/latex-ocr works
pix2tex is a LaTeX OCR system: users submit or capture a formula image, the model preprocesses it and generates LaTeX, which an interface returns or displays. The repository also includes a dataset-building and model-training workflow, plus evaluation. The graph preserves documented product capabilities even where implementation excerpts were unavailable; edges for those relationships cite the README, and unsampled internals are represented only at paths supported by the tree.
Main entry points: cli.py LatexOCR · utils.py Model assembly · arxiv.py arXiv ingestion · gui.py Desktop GUI
Components 26
User interfaces
- CLI
pix2tex/cli.py - Desktop GUI
pix2tex/gui.py - Prediction API
pix2tex/api/app.py - Streamlit demo
pix2tex/api/streamlit.py - MathJax rendering
pix2tex/resources/MathJax.js
OCR inference
- LatexOCR
pix2tex/cli.py - Image transforms
pix2tex/dataset/transforms.py - Resolution model
pix2tex/train_resizer.py - Checkpoint download
pix2tex/model/checkpoints/get_latest_checkpoint.py - Model assembly
pix2tex/models/utils.py - Hybrid ViT encoder
pix2tex/models/hybrid.py - ViT encoder
pix2tex/models/vit.py - Transformer decoder
pix2tex/models/transformer.py
Dataset preparation
- Paired dataset
pix2tex/dataset/dataset.py - Formula scraping
pix2tex/dataset/scraping.py - arXiv ingestion
pix2tex/dataset/arxiv.py - LaTeX extraction
pix2tex/dataset/extract_latex.py - Macro expansion
pix2tex/dataset/demacro.py - Formula rendering
pix2tex/dataset/render.py - PDF to PNG
pix2tex/dataset/latex2png.py - Formula preprocessing
pix2tex/dataset/preprocessing/preprocess_formulas.py - Vocabulary generation
pix2tex/dataset/preprocessing/generate_latex_vocab.py
Model development
- Model training
pix2tex/train.py - Model evaluation
pix2tex/eval.py
Other components
- Image user
- Training data
Connections 26
- Image user → CLI: submits image
- Image user → Desktop GUI: captures image
- Image user → Prediction API: uploads image
- CLI → LatexOCR: invokes
pix2tex/__main__.py - Desktop GUI → LatexOCR: predicts with
pix2tex/gui.py - Prediction API → LatexOCR: predicts with
pix2tex/api/app.py - Streamlit demo → Prediction API: calls API
- LatexOCR → Image transforms: transforms image
pix2tex/cli.py - LatexOCR → Resolution model: resizes image
- LatexOCR → Model assembly: loads model
pix2tex/cli.py - LatexOCR → Checkpoint download: loads weights
pix2tex/cli.py - Model assembly → Hybrid ViT encoder: builds encoder
pix2tex/models/utils.py - Model assembly → ViT encoder: builds encoder
pix2tex/models/utils.py - Model assembly → Transformer decoder: builds decoder
pix2tex/models/utils.py - Desktop GUI → MathJax rendering: renders output
- Training data → Paired dataset: pairs labels
- Formula scraping → LaTeX extraction: extracts formulas
pix2tex/dataset/scraping.py - arXiv ingestion → LaTeX extraction: extracts formulas
pix2tex/dataset/arxiv.py - arXiv ingestion → Formula scraping: uses scraping
pix2tex/dataset/arxiv.py - arXiv ingestion → Macro expansion: expands macros
pix2tex/dataset/arxiv.py - Macro expansion → LaTeX extraction: removes labels
pix2tex/dataset/demacro.py - Formula rendering → PDF to PNG: converts PDF
pix2tex/dataset/render.py - Formula preprocessing → Vocabulary generation: supports tokenization
- Paired dataset → Model training: supplies batches
pix2tex/train.py - Model training → Model evaluation: evaluates model
pix2tex/train.py - Paired dataset → Model evaluation: supplies examples
pix2tex/eval.py