CannyEdit: Selective Canny Control and Dual-Prompt Guidance for Training-free Image Editing

Recent text-to-image advances enable training-free regional image editing by leveraging generative priors of foundation models. However, existing methods struggle to balance text adherence in edited regions, context fidelity in unedited areas, and seamless integration of edits. We introduce CannyEdit, a novel training-free framework addressing this trilemma through two key innovations. First, Selective Canny Control uses structural guidance from a Canny ControlNet exclusively on unedited regions, preserving original details while enabling precise, text-driven edits. Second, Dual-Prompt Guidance combines a local prompt for specific edits with a global prompt for overall coherence. This synergy achieves controllable local editing for object addition, replacement, and removal, offering a superior trade-off among text adherence, context fidelity, and seamlessness compared to mask-based methods. Moreover, CannyEdit works effectively with rough masks or single-point hints for addition tasks and integrates seamlessly with vision-language models without training. Extensive evaluations highlight CannyEdit’s strong performance in complex instruction-based editing and object addition scenarios.