Penquify

•
7
•

Generates photorealistic smartphone photos of logistics documents with induced defects and verified ground truth.

Penquify is a Python toolkit designed to create high-fidelity synthetic datasets for training and testing Document AI and OCR models. It acts as an 'OCR in reverse' by intentionally degrading documents, simulating real-world imperfections like coffee stains, folds, blur, and skew. Starting from simple data inputs like JSON or existing PDFs, it generates clean, realistic logistics documents (e.g., dispatch guides), then transforms them into photorealistic smartphone images, complete with verified ground truth for every data field and an occlusion manifest. This allows developers and data scientists to generate vast, diverse datasets for robust model training without needing real-world document collection.

主な機能

01Generates photorealistic smartphone photos of logistics documents with various defects (e.g., blur, folds, stains).
02Produces verified ground truth for every field and an occlusion manifest for each generated photo variation.
03Automates realistic document generation (PDF) from JSON, existing PDFs, or natural language descriptions.
04Provides multiple interfaces: CLI, Python API, FastAPI REST API, MCP Server tools, and Agent SDK plugin.
05Supports natural language prompts and a detailed schema for configuring photo variations and camera presets.
063 GitHub stars

ユースケース

01Creating large-scale synthetic datasets to train and evaluate Document AI and OCR models.
02Testing the robustness and resilience of document processing systems against real-world imperfections and data discrepancies.
03Rapidly prototyping and expanding machine learning models for logistics document recognition and extraction.