FableBench: A Structured Fable Dataset for Literary Story Visualization

In current datasets such as PororoSV and FlintstonesSV, most of them are fixed animation segments with relatively few characters, short sentences, and consistent visual styles. Therefore, relying solely on these datasets to evaluate story visualization models may underestimate the difficulty brought by open literary texts.

We have built a new dataset FableBench to address the above issues. This dataset selected 426 publicly available fables, and obtained 2788 scene level micro scenes through text collection, cleaning, narrative decomposition, prompt word organization, character visual archive extraction, and quality inspection. Each story contains a cleaned main text, symbolic tags, ordered micro scenes, Chinese scene descriptions, English image prompts, and structured character information.