Scholarlink

Recent advances in text-to-image (T2I) generation have made substantial progress in generating images from textual prompts. However, they frequently fail to produce image stories with rich plot structures, which is essential in fields of creative generation. In this paper, we introduce StoryBench, a diverse, explainable, multi-hop evaluation dataset for narrative T2I generation, comprising 728 prompts across five categories: animal-nature, labor, medical, sport and technology. We assess 5 prominent T2I models, including proprietary models DALL \(\cdot \) E 3 and Midjourney, and found that even advanced T2I models exhibit limited capability in generating complex event-narrative images.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

StoryBench: A Dataset for Diverse, Explainable, Multi-hop Narrative Text-to-Image Generation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yuan Ge,
  • Kaiyang Ye,
  • Saihan Chen,
  • Aokai Hao,
  • Xiangnan Ma,
  • Kaiyan Chang,
  • Tong Xiao,
  • Jingbo Zhu

摘要

Recent advances in text-to-image (T2I) generation have made substantial progress in generating images from textual prompts. However, they frequently fail to produce image stories with rich plot structures, which is essential in fields of creative generation. In this paper, we introduce StoryBench, a diverse, explainable, multi-hop evaluation dataset for narrative T2I generation, comprising 728 prompts across five categories: animal-nature, labor, medical, sport and technology. We assess 5 prominent T2I models, including proprietary models DALL \(\cdot \) E 3 and Midjourney, and found that even advanced T2I models exhibit limited capability in generating complex event-narrative images.