Bounded Editing: Multi-Object Image Manipulation with Region-Specific Control

Abstract

Recent diffusion-based models have achieved significant success in vision domains such as image generation, and text-guided image manipulation. Text-guided image editing aims for users to modify specific objects and their attributes based on textual descriptions. However, current image editing approaches are susceptible to unintended modifications to non-target regions or other target regions in images when altering multiple objects. Some studies depend on detailed masks that are challenging to obtain for fine-grained image editing. To address these issues, we propose Bounded Editing, which allows for precise manipulation of specific areas. First, our approach separates the target and non-target regions from the image using bounding boxes. Second, we propose a guidance loss that enhances editing capabilities, enabling precise modifications to target objects while preventing undesired changes to background. By integrating our method with an existing image editing framework, we achieve significant improvements over state-of-the-art methods. Extensive experiments demonstrate the effectiveness of our proposed approach in changing objects, and modifying attributes such as colors and materials, especially on multi-object editing scenarios.