![[2602.21824] DocDjinn: Controllable Synthetic Document Generation with VLMs and Handwriting Diffusion](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.21824] DocDjinn: Controllable Synthetic Document Generation with VLMs and Handwriting Diffusion
Summary
DocDjinn introduces a framework for generating synthetic documents using Vision-Language Models (VLMs), addressing challenges in data acquisition for document intelligence tasks.
Why It Matters
The ability to generate high-quality synthetic documents can significantly reduce the costs and labor associated with data annotation, while also preserving privacy. This innovation could enhance various document understanding applications, making it a crucial development in machine learning and AI.
Key Takeaways
- DocDjinn leverages VLMs for controllable synthetic document generation.
- The framework can produce annotated documents from just 100 real training samples.
- Synthetic documents maintain semantic consistency and visual plausibility.
- The approach addresses privacy concerns in data acquisition.
- 140k+ synthetic document samples and code are publicly available for further research.
Computer Science > Machine Learning arXiv:2602.21824 (cs) [Submitted on 25 Feb 2026] Title:DocDjinn: Controllable Synthetic Document Generation with VLMs and Handwriting Diffusion Authors:Marcel Lamott, Saifullah Saifullah, Nauman Riaz, Yves-Noel Weweler, Tobias Alt-Veit, Ahmad Sarmad Ali, Muhammad Armaghan Shakir, Adrian Kalwa, Momina Moetesum, Andreas Dengel, Sheraz Ahmed, Faisal Shafait, Ulrich Schwanecke, Adrian Ulges View a PDF of the paper titled DocDjinn: Controllable Synthetic Document Generation with VLMs and Handwriting Diffusion, by Marcel Lamott and 13 other authors View PDF Abstract:Effective document intelligence models rely on large amounts of annotated training data. However, procuring sufficient and high-quality data poses significant challenges due to the labor-intensive and costly nature of data acquisition. Additionally, leveraging language models to annotate real documents raises concerns about data privacy. Synthetic document generation has emerged as a promising, privacy-preserving alternative. We propose DocDjinn, a novel framework for controllable synthetic document generation using Vision-Language Models (VLMs) that produces annotated documents from unlabeled seed samples. Our approach generates visually plausible and semantically consistent synthetic documents that follow the distribution of an existing source dataset through clustering-based seed selection with parametrized sampling. By enriching documents with realistic diffusion-based handwrit...