WAS-R, WAS-S

The WAS dataset was jointly created by Huazhong University of Science and Technology and Adobe, focusing on artistic text segmentation tasks. The WAS‑R dataset contains 7,100 artistic text images with high‑quality word‑level annotations, including quadrilateral boxes, masks, and transcripts. The WAS‑S dataset is a synthetic set generated using large multimodal and diffusion models, aimed at improving accuracy and generalization of text‑segmentation models. These datasets are primarily used to address artistic‑text segmentation in complex scenes, especially for text‑image generation, editing, and style‑transfer tasks.