MuST-Bench

该数据集旨在评估视觉文本生成模型在不同书写系统间进行翻译时，能否保持设计意图的能力。该数据集专门用于评估模型在生成英文、中文和韩文排版方面的表现，并支持多模态风格翻译。数据集包含了每种语言1,040个风格化字符，这些字符源自10个不同的风格主题，每个语言有26个提示。该任务为多模态风格翻译。

This dataset is designed to evaluate the capability of visual text generation models to preserve design intent when translating between different writing systems. It is specifically developed to assess the performance of models in generating English, Chinese and Korean typography, and supports multimodal style translation. The dataset includes 1,040 stylized characters for each language, which are derived from 10 distinct style themes, with 26 prompts provided per language. This task is multimodal style translation.