Openstory++ Large-Scale Image Instance Dataset

The Openstory++ dataset was jointly developed by South China University of Technology, Westlake University, OPPO US Research Center, and King Abdullah University of Science and Technology in 2024.

Openstory++ is designed to solve the problem that existing image generation models have difficulty maintaining instance consistency in long text contexts. It provides a rich resource by combining instance-level annotations of images and text, enabling it to generate highly consistent images in long text contexts. The development of this dataset is based on a deep understanding of the problem that existing image generation models lack consistency when dealing with complex narratives. Through automated keyframe extraction, subtitle generation by visual-language models, and narrative coherence polishing by large language models, a large-scale resource library that supports complex narrative generation tasks is constructed.