datajuicer/Img-Diff

Img-Diff是一个专注于为多模态大型语言模型描述物体差异的高质量合成数据集。它通过对比学习原理，使用Stable-Diffusion-XL模型和高级图像编辑技术创建的物体替换样本对，来增强模型在识别相似图像中匹配和不同部分的能力。数据集包含了一个用于识别物体差异的区域生成器和用于生成详细差异描述的差异标注生成器。

Img-Diff is a high-quality synthesis dataset focusing on describing object differences for Multimodal Large Language Models (MLLMs). It enhances the models ability to identify matching and distinct components in similar images by leveraging contrastive learning principles, utilizing the Stable-Diffusion-XL model, and advanced image editing techniques to create pairs of images highlighting object replacements. The dataset includes a Difference Area Generator for identifying object differences and a Difference Captions Generator for detailed descriptions of the differences.