Stepwise Attention Control for Training-Free Image Editing

Text-conditioned image editing aims to modify visual content based on natural language prompts while preserving the spatial structure of the input image. This capability plays a vital role in applications such as creative media design, virtual scene generation, and interactive image manipulation. However, achieving both semantic fidelity and structural preservation remains a major challenge in diffusion-based approaches. Many existing methods exhibit an imbalance between preserving the input image’s structure and reflecting the semantic intent of the prompt, often resulting in suboptimal editing outcomes. To address this problem, we propose a training-free method called Stepwise Attention Control, which divides the denoising process into four stages based on the diffusion time step and applies tailored attention control strategies to each stage. This stage-wise approach enables balanced control of both the structural fidelity and semantic consistency during the editing process. Experimental results demonstrate that the proposed method achieves more faithful semantic edits while more effectively retaining the structural features of the original image compared to existing techniques.