Disentangling Writer and Character Styles for Handwriting Generation
Abstract: Training machines for synthesizing diverse handwritings is an intriguing task. Recently, some RNN-based methods are proposed to generate stylized online Chinese characters. But these methods mainly focus on learning a person’s overall writing style and hence neglect the detailed style inconsistencies between characters from the same writer. For example, one person’s handwritings always appear an overall uniformity (e.g., character slant and aspect ratios) but there are still small style differences between local regions (e.g., stroke length and curvature) of characters. Motivated by this, in this paper, we propose to disentangle the style representations at both writer and character levels from individual handwritings. Specifically, we propose the style-disentangled transformer (SDT), equipped with two complementary contrastive objectives, to extract the overall writer-wise and detailed character-wise style representations, respectively, which boosts the generation quality of online handwritings. Extensive experiments on various language scripts verify the superiority of SDT. Particularly, we empirically find that the two learned style representations provide information with different frequency magnitudes, which demonstrates the necessity of separate style extraction.
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.
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: Yes
Please Choose The Closest Area That Your Submission Falls Into: Generative models
Community Implementations: [ 4 code implementations](https://www.catalyzex.com/paper/disentangling-writer-and-character-styles-for/code)
5 Replies
Loading