ConceptEdit-12M
收藏资源简介:
ConceptEdit-12M是由上海交通大学和蚂蚁集团联合构建的大规模高质量图像编辑数据集,包含1200万对精心验证的编辑样本。该数据集基于超过1000个细粒度编辑概念的分层分类法,通过改进的合成框架生成,利用大语言模型蒸馏世界知识以覆盖广泛场景,并采用VQA过滤确保高保真度。其核心创新在于突破传统依赖源图像多样性的范式,转而强调编辑概念的丰富性和粒度,同时引入密集监督策略提升训练效率。该数据集旨在解决图像编辑模型泛化能力不足和训练稀疏性问题,为指令式图像编辑提供更全面的训练与评估基准。
ConceptEdit-12M is a large-scale high-quality image editing dataset jointly constructed by Shanghai Jiao Tong University and Ant Group, comprising 12 million carefully validated editing sample pairs. This dataset is built upon a hierarchical taxonomy of over 1,000 fine-grained editing concepts, generated via an improved synthesis framework that distills world knowledge using Large Language Models (LLMs) to cover diverse scenarios, and employs VQA-based filtering to guarantee high fidelity. Its core innovation lies in breaking through the traditional paradigm that relies on the diversity of source images, instead emphasizing the richness and granularity of editing concepts, while introducing dense supervision strategies to enhance training efficiency. This dataset aims to address the issues of insufficient generalization capability and training sparsity of image editing models, providing a more comprehensive training and evaluation benchmark for instruction-guided image editing.

- 1Unlocking the Potential of Image Editing via Concept Scaling and Dense Supervision上海交通大学; 蚂蚁集团 · 2026年



