ViStoryBench

ViStoryBench是一个全面的故事可视化评估基准，旨在推动故事可视化框架在各种实际场景中的应用。数据集包含80个故事片段，共有344个角色和509个参考图像，涵盖了多种故事类型和艺术风格。该数据集经过精心策划，以平衡叙事结构和视觉元素，并包括具有单一和多个主角的故事，以测试模型在保持角色一致性方面的能力。此外，它还包括复杂的情节和精细的世界构建，以挑战模型生成准确视觉内容的能力。ViStoryBench通过广泛的评估指标，包括自动评估指标，为研究人员提供了一个结构化和多方面的框架，以全面识别不同模型的优势和劣势，从而促进有针对性的改进。

ViStoryBench is a comprehensive story visualization evaluation benchmark developed to advance the deployment of story visualization frameworks across diverse real-world scenarios. The dataset consists of 80 story segments, with 344 distinct characters and 509 reference images in total, covering a wide spectrum of story genres and artistic styles. It has been carefully curated to strike a balance between narrative structure and visual elements, and includes stories featuring both single and multiple protagonists to evaluate models' capacity to sustain character consistency. Additionally, it integrates complex plotlines and elaborate worldbuilding to test models' ability to generate visually accurate content. ViStoryBench offers researchers a structured, multi-faceted evaluation framework via a broad range of metrics, including automatic evaluation metrics, to comprehensively discern the strengths and weaknesses of different models, thus facilitating targeted model improvements.