Multi-Edit Bench
收藏资源简介:
Multi-Edit Bench是一个新的基准数据集,旨在评估迭代图像编辑能力。该数据集通过引入层状语义评估指标和交互式编辑场景,以评估图像编辑方法的编辑质量和跨修改一致性。数据集的设计考虑了实际编辑工作流程,允许用户通过添加、删除或重新定位对象来逐步细化场景,从而测量方法在多次、可能复杂的修改中保持上下文和保持一致构图的能力。
Multi-Edit Bench is a novel benchmark dataset developed to evaluate iterative image editing capabilities. This dataset introduces layered semantic evaluation metrics and interactive editing scenarios to assess the editing quality and cross-modification consistency of image editing approaches. Designed with real-world editing workflows in mind, the dataset enables users to gradually refine scenes by adding, deleting, or relocating objects, thereby measuring the ability of editing methods to retain contextual coherence and consistent composition across multiple, potentially complex modifications.

- 1Improving Editability in Image Generation with Layer-wise Memory首尔国立大学 · 2025年



