# Visual Text Generation in the Wild

**Authors:** Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, Zhibo Yang  
**arXiv:** [2407.14138](https://arxiv.org/abs/2407.14138)  
**Date:** 2024-07-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.14138v1/

## Key points

- Authors: Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, Zhibo Yang
- Topic: Visual Text Generation in the Wild
- Challenges faced in rendering high-quality text images in real-world scenarios
- Introduction of SceneVTG as a novel visual text generator to address limitations
- SceneVTG leverages Multimodal Large Language Model for recommending text regions and contents
- Use of conditional diffusion model with recommendations to generate high-quality text images
- Performance superiority of SceneVTG over traditional rendering-based and recent diffusion-based methods demonstrated through experiments
- Superior utility of generated text images for tasks like text detection and recognition
- Availability of code and datasets at AdvancedLiterateMachinery for enhanced accessibility and reproducibility

## Summary

In their paper titled "Visual Text Generation in the Wild," authors Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, and Zhibo Yang discuss recent advancements in generative models that have led to significant progress in the field of visual text generation. The authors highlight the challenges faced in rendering high-quality text images in real-world scenarios and propose a novel visual text generator called SceneVTG to address these limitations. <br>
SceneVTG leverages a Multimodal Large Language Model to recommend reasonable text regions and contents across various scales and levels. These recommendations are then used by a conditional diffusion model as conditions to generate high-quality text images. Extensive experiments demonstrate that SceneVTG outperforms traditional rendering-based methods and recent diffusion-based methods in terms of fidelity and reasonability. Additionally, the generated text images offer superior utility for tasks such as text detection and recognition. <br>
The availability of code and datasets at AdvancedLiterateMachinery further enhances the accessibility and reproducibility of their research findings. Overall, by introducing an innovative approach that addresses key challenges faced in rendering high-quality text images in real-world scenarios.