ReChar: Revitalising Characters with Structure-Preserved and User-Specified Aesthetic Enhancements

Zhongyu Yang1, Zuhao Yang2, Zhang Luo3, Yonghan Zhang4, Yang Xu5, Wei Pang4, Jingfen Lan6, Yingfang Yuan4,†
1School of Mathematics and Statistics, Lanzhou University
2School of Computer Science and Engineering, Nanyang Technological University
3SeaFog AI
4School of Mathematical and Computer Sciences, Heriot-Watt University
5Department of Computer Science and Engineering, Southern University of Science and Technology
6School of Mathematics and Statistics, Xidian University
†Corresponding author
Paper Dataset(Coming Soon) Code(Coming Soon)
MY ALT TEXT

Given a variety of text prompts, ReChar can generate images that adhere to both style image and text instruction without test time fine-tuning, which demonstrates remarkable consistency in stylised character generation. Best viewed when zoomed in.

Abstract

AI-Generated Content (AIGC) has recently surged in popularity, driven by its efficiency, consistent output quality, and versatile customization capabilities. While applications of image generation extend across numerous domains, research into stylized language writing system generation remains limited, particularly within the context of aesthetic enhancement. Inspired by the pictogram subset of Chinese characters, we propose ReChar, which aims to create visually appealing and culturally rich artworks that seamlessly integrate text-guided decorative elements and stylistic preferences within character structures. Additionally, we introduce ImageNet-ReChar, the benchmark dataset designed to assess stylized character generation, with a focus on the generalization of user-defined elements.Our evaluation across multiple languages (Chinese, Japanese, English, and Latin) demonstrates ReChar’s ability to preserve original character structures while embedding decorative elements into the strokes and incorporating user-defined styles to achieve aesthetic enhancement.Our qualitative and quantitative analyses reveal that ReChar consistently outperforms existing methods.

Method

MY ALT TEXT

Our ReChar Framework. ReChar integrates three distinct yet interrelated modules: (1) a character structure extraction module, which is designed to preserve the integrity of the character's form, (2) an element generation module, responsible for producing user-defined decorative elements based on textual input, and (3) a style extraction module, aimed at capturing the visual style from a reference image provided by the user. These components are subsequently fused in a controllable synthesis step, which enables flexible and user-customized image generation. To provide a clearer understanding of our approach, we will illustrate the generation process of an instance through a detailed case study.

✨ Qualitative Results ✨

MY ALT TEXT

Visual comparisons between our proposed ReChar with other existing methods, conditioned on different contents and styles. For existing methods, we used their official implementations and default settings, adjusting them only when necessary. To enhance the visual representation of the results, the edges of the characters structure are highlighted in pink.Best viewed when zoomed in.

✨ Ablation Study ✨

MY ALT TEXT

The effect of each strategy. We regard ReChar as the baseline.w/o denotes the absence of this component.Notably, our method achieves the best visual effect. To enhance the visual representation of the results, the edges of the characters structure are highlighted in pink.

✨More Result ✨

BibTeX

BibTex Code Here