What changed
Hugging Face introduced a new open-source tool called Recapture, which allows data engineers to generate realistic scans and photos from clean digital documents. This tool is designed to help diversify AI training sets by creating synthetic document images that mimic real-world conditions.
What this means in practice
Recapture takes a clean page image and its layout boxes, then renders it as a realistic phone photo or flatbed scan. Users only need to label the page once, and Recapture produces multiple degraded copies with the corresponding labels already applied. This significantly reduces the effort required to create diverse document datasets for AI training.
Limitations and availability
Recapture is an open-source tool developed by Nutrient and is available on GitHub. It requires Python 3.12 and uv for installation. The tool supports both scan mode and photo mode, with scan mode being faster and not requiring a GPU. Photo mode involves more complex processing, including 3D augmentation to simulate real-world document conditions. The tool also integrates with Nutrient Parse, an API for document parsing, to streamline the dataset creation process. The tool is intended for use by data engineers and developers looking to enhance their AI training datasets with realistic document variations.
