TextSculptor: Training and Benchmarking Scene Text Editing
Published in arxiv , 2026
Yiheng Lin, Siyu Jiao, Xiaohan Lan, Wei Zhou, Qi She, Fei Yu, Heyun Chen, Zhengwei Wang, Jinghuan Chen, Moran Li, Yingchen Yu, Zijian Feng, Yao Zhao, Yunchao Wei, Yujie Zhong
Paper · Code · Data · Benchmark
TextSculptor is a comprehensive framework for data construction and evaluation in scene text editing. It introduces TextSculpt-Data, a dataset with 3.2 million training samples: 1.2 million OCR-verified text-to-image samples and 2 million paired text-editing samples with aligned source and target images.
The work also presents TextSculpt-Bench, covering text addition, replacement, removal, and hybrid editing. Its evaluation protocol jointly measures text accuracy, visual quality, and background preservation, providing a standardized way to assess scene-text editing systems.