Openstory++: A Large-scale Dataset and Benchmark for Instance-aware Open-domain Visual Storytelling

We introduce OpenStory++, a large-scale open-domain dataset contains focusing on enabling MLLMs to perform storytelling generation tasks.

related resorcce

paper: https://arxiv.org/abs/2408.03695

code: https://github.com/YeLuoSuiYou/openstorypp

News

2024/7/31 We have reorganized and distributed the high-quality subset and released most of the story data collected from YouTube. Due to copyright issues, we have not released the raw images, but we will provide the method of organizing the dataset later.

2024/6/22 we release the high-quality subset of our dataset’s unique data, which contains about 15M sample.

Download Method

1. We have replaced the original image with a URL, you can use img2dataset to download single image dataset.
2. You can find the story dataset organization methods in our code repositories, where we provide video download and dataset reorganization methods.