Open document AI tooling, training data and vision-language models

Nutrient open-sources Recapture to turn clean pages into labeled document photos

The Apache-2.0 tool uses simulated scans and 3D phone photography to expand OCR and document-layout training sets without relabeling each image.

Open Model Weights published20 Nov 2026
Primary sourceNutrient engineering team — original Recapture technical article on Hugging Face
Source published2026-11-18
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original minimalist editorial illustration representing vision in AI research; conceptual, not actual model output.
AI-generated editorial illustration · Not official model imagery.

Open document-understanding models often learn from pages that look nothing like the images they encounter in production. Training sets may contain clean PDF renders and idealized layouts, while real users photograph folded invoices in poor light or scan aging documents with streaks and skew. Nutrient has released Recapture, an Apache-2.0 open-source tool designed to make that mismatch easier to study and address. Primary source ↗

Recapture takes a clean page image with spatial labels and generates visually degraded variants while transforming the annotation coordinates with the page. The goal is more than applying random blur. If a paragraph, table or signature moves under a simulated camera angle or paper fold, the matching label should move as well. This matters for layout detection, OCR pipelines and other document AI systems that need dependable region-level supervision.

The project has two distinct rendering paths. Scan mode synthesizes effects such as low toner, ink spread, dust, dark edges and imperfect paper alignment. It does not require a GPU. Photo mode builds a deformed page mesh, places it against a virtual desktop, adds lighting and shadows, and renders a phone-camera image with lens and compression effects. Nutrient documents different desk textures, object props and camera-quality controls so one labeled source can produce diverse training examples.

The exported data is intended to remain usable rather than just visually impressive. Recapture supports COCO and YOLO-OBB formats, carries per-image generation settings, and keeps variants of the same source page in the same dataset split. That grouping helps prevent near-identical images from leaking from training into validation or testing. Its rendering configuration can also be seeded to reproduce a specific synthetic sample.

Nutrient's official repository provides a local Python workflow and configuration-based rendering. Existing label files can be used without a commercial service; the team's separate Nutrient Parse API is an optional way to generate spatial labels for documents that do not yet have them. The open-source repository does not make Recapture itself a new model checkpoint, and it should not be confused with a released OCR foundation model.

The larger significance is methodological. Synthetic data can systematically probe how open-weight document models behave under folds, blur, occlusion and scanning artifacts. But realism and volume alone do not guarantee better results on authentic business documents. Researchers should compare against held-out real scans, audit label fidelity and test whether synthetic scenes introduce shortcuts. Recapture creates a practical, reproducible tool for those experiments, not an automatic claim of improved model accuracy.

OMW REGISTRY WATCH

What this changes for the evidence layer

Recapture is an open-source dataset augmentation tool, not new open-weight model weights. The code repository declares Apache-2.0. Claims that augmented sets improve downstream model accuracy require separate controlled experiments; no OMW benchmark is implied. Optional Nutrient Parse is a separate API and is not required if layout labels already exist.

PRIMARY SOURCE

Nutrient engineering team — original Recapture technical article on Hugging Face

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗