JourneyDB
收藏资源简介:
JourneyDB是由香港中文大学多媒体实验室创建的一个大规模数据集,专注于生成图像的多模态视觉理解。该数据集包含400万张高质量生成图像,每张图像都配有用于生成该图像的文本提示。此外,JourneyDB还引入了一个包含22种不同文本到图像生成模型的外部子集,使其成为评估生成图像理解能力的全面基准。数据集创建过程中,通过Discord平台收集图像,并利用GPT-3.5模型进行详细的下游任务标注,包括风格和内容的分离、图像标题生成以及视觉问答等。JourneyDB的应用领域主要集中在提升生成内容理解的研究,特别是在内容和风格的细粒度识别上,旨在解决现有模型在处理生成数据时遇到的挑战。
JourneyDB is a large-scale dataset created by the Multimedia Laboratory of The Chinese University of Hong Kong, focusing on multimodal visual understanding of generated images. This dataset contains 4 million high-quality generated images, each paired with the text prompt used to generate it. Additionally, JourneyDB includes an external subset consisting of 22 distinct text-to-image generation models, making it a comprehensive benchmark for evaluating generated image understanding capabilities. During the dataset construction process, images were collected via the Discord platform, and detailed downstream task annotations were performed using the GPT-3.5 model, including style-content separation, image caption generation, Visual Question Answering (VQA), and more. The application scenarios of JourneyDB primarily focus on research for improving generated content understanding, especially fine-grained recognition of content and style, aiming to address the challenges encountered by existing models when processing generated data.

- 1JourneyDB: A Benchmark for Generative Image Understanding多媒体实验室,香港中文大学 · 2023年



