DEAL-300K
收藏资源简介:
DEAL-300K是一个专注于扩散基图像编辑区域定位(DIML)的大规模数据集,包含超过30万张经过标注的图像。该数据集旨在应对扩散基图像编辑技术带来的挑战,这些技术能够实现与原始内容无缝融合的语义级编辑。DEAL-300K的构建利用了多模态大型语言模型(MLLMs)进行指令驱动的编辑,并结合主动学习标注流程,确保了数据的多样性和高质量。数据集中的原始图像来源于MS COCO数据集,编辑指令由细调的Qwen-VL生成。DEAL-300K支持多种图像尺寸(从128×512到512×576不等),适用于通用场景的扩散基图像编辑定位任务。与其他公开的DIML数据集相比,DEAL-300K在规模和多样性上具有显著优势。
DEAL-300K is a large-scale dataset focused on diffusion-based image editing region localization (DIML), containing over 300,000 annotated images. This dataset aims to address the challenges posed by diffusion-based image editing techniques, which enable semantic-level edits that seamlessly fuse with original image content. The construction of DEAL-300K leverages multimodal large language models (MLLMs) for instruction-driven editing, paired with an active learning annotation workflow to ensure data diversity and high quality. The original images in the dataset are sourced from the MS COCO dataset, while the editing instructions are generated by fine-tuned Qwen-VL. DEAL-300K supports multiple image sizes ranging from 128×512 to 512×576, and is applicable to diffusion-based image editing localization tasks for general scenarios. Compared with other publicly available DIML datasets, DEAL-300K exhibits significant advantages in terms of scale and diversity.




