Abstract

Despite the recent advances in layer decomposition, instruction-guided layer editing remains a pivotal yet under-explored challenge in image manipulation. Existing solutions typically rely on decoupled cascaded scheme: compositing layer onto solid background, applying off-the-shelf RGB editor, and then re-matting foregrounds. This way inherently neglects global contextual harmony across layers and struggles to preserve intricate boundary details during matting. To address this, we present LayerEdit, an end-to-end framework for context-aware image layer editing built upon a pre-trained layer decomposition foundation model. Unlike the decoupled cascaded scheme, LayerEdit learns joint layer editing and high-fidelity matting by seamlessly synergizing editing priors from massive RGB data with the intrinsic hierarchical perception embedded in the decomposition backbone. Furthermore, rather than processing the input layer in isolation, LayerEdit amplifies context awareness by integrating the global semantic guidance from the holistic composite into the input layer. Empowered by the introduced RGB-to-RGBA knowledge transfer strategy and the effective utilization of the native RGBA structural representations, our model achieves exceptional layer manipulation within a single network relying on limited curated layer editing data. Extensive experiments demonstrate that LayerEdit outperforms cascaded baselines.