UDP-Edit: Union Dual-Prompt Attention for Local Image Editing in Fast Diffusion Models

Fast-sampling diffusion models have become prominent in text-to-image editing due to their significantly reduced inference time. However, achieving precise local edits with these models remains challenging. For example, altering a teddy bear to a wooden bear should affect only the object, not the background. Prior approaches fail to balance precision and efficiency, often resulting in either global changes or slow inference. In this paper, we propose UDP-Edit, a union dual-prompt cross-attention editing framework for controlled local editing under fast diffusion sampling. UDP-Edit merges cross-attention maps from both source and target prompts to generate a soft semantic mask. This mask constrains edits to relevant regions, preserving global structure while enabling localized transformation. Experiments demonstrate that UDP-Edit achieves accurate local edits and preserves sampling efficiency, validated through both qualitative and quantitative metrics.