DenseFusion-1M

该数据集包含了来自未筛选的LAION数据集的100万张具有高度代表性的图片，并通过一款标题生成引擎为这些图片生成了密集的描述，以增强多模态感知能力。该数据集旨在提升现有多语言大模型在多样化的视觉-语言基准测试中的感知和认知能力，特别是针对高分辨率图像。规模达到了100万张图片，任务聚焦于多模态感知和视觉-语言理解。This dataset contains 1 million highly representative images sourced from the unfiltered LAION dataset. Dense captions were generated for these images using a caption generation engine to enhance multimodal perception capabilities. It is designed to improve the perceptual and cognitive abilities of existing multilingual large language models (LLMs) across diverse vision-language benchmarks, particularly for high-resolution images. With a total of 1 million images, this dataset focuses on multimodal perception and vision-language understanding tasks.