Early Semantic Grounding in Image Editing Models for Zero-Shot Referring Image Segmentation
The work investigates if image editing models inherently possess visual semantic grounding, which is then used for zero-shot referring image segmentation (RIS). The researchers found that strong foreground-background separability exists in these models' internal representations at the earliest denoising step. They propose a training-free framework that uses this inherent separability to decompose localization into spatial attention and semantic discrimination, achieving superior performance.