Write More at Once: Stylized Chinese Handwriting Generation via Two-stage Diffusion

Honglie Wang, Minsi Ren, Yangyang Liu, Yan-Ming Zhang

Keywords: Handwritten Text Generation; Conditional Diffusion;

Abstract: Handwritten data generation is an intriguing research area with broad applications in human interaction with digital documents. In Chinese handwritten text generation, practical applications necessitate the ability to produce sentence-level handwritten data to convey complex information effectively. However, existing methods mainly focus on generating single-font outputs. To tackle this challenge, we model handwritten text generation as a \textit{style transfer problem}, aiming to convert a standard text line template into a target handwriting style. Recognizing the highly structured nature of handwritten data, we view complex text lines as compositions of individual characters and their positions. We propose a two-stage text line generation method based on generative diffusion model. In the first stage, character positions are generated using a Character-Position-Diffusion (CharPos-Diff), which, combined with standard character templates from a digital library, creates text line-level templates. In the second stage, a font style transfer diffusion model (Imitating-Diff) generates handwritten text lines directly from these templates. Our extensive experiments show that our method effectively mimics handwriting styles, generates structurally accurate text lines, and facilitates the simultaneous generation of paragraph-level handwritten text.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9869