RefEdit-Data
收藏资源简介:
RefEdit-Data是一个基于指示的图像编辑数据集,由亚利桑那州立大学创建。该数据集包含超过20000个图像编辑三元组,旨在帮助图像编辑模型在复杂的场景中准确识别和编辑多个实体。数据集通过一个可扩展的合成数据生成管道创建,使用GPT-4o生成复杂的编辑任务,并整合了Grounded Segment Anything和FlowChef技术进行精确、受控的编辑。RefEdit-Data被用于训练RefEdit模型,该模型在RefEdit-Bench基准测试中表现出色,超越了在数百万数据上训练的现有方法。
RefEdit-Data is an instruction-based image editing dataset created by Arizona State University. This dataset contains over 20,000 image editing triplets, aiming to assist image editing models in accurately identifying and editing multiple entities in complex scenarios. It is constructed via a scalable synthetic data generation pipeline, where GPT-4o is used to generate complex editing tasks, and Grounded Segment Anything and FlowChef technologies are integrated to enable precise and controlled editing. RefEdit-Data is utilized to train the RefEdit model, which achieves excellent performance on the RefEdit-Bench benchmark and outperforms existing methods trained on millions of data samples.




