DiffEditor is a novel method designed to enhance the accuracy and flexibility of diffusion-based image editing. It addresses two key issues: (1) lack of editing accuracy and unexpected artifacts in complex scenarios, and (2) insufficient flexibility in harmonizing editing operations. The method introduces image prompts to fine-grained image editing, which, when combined with text prompts, provide more detailed descriptions of the editing content. To improve flexibility while maintaining content consistency, DiffEditor combines stochastic differential equations (SDE) with ordinary differential equations (ODE) sampling. Additionally, it incorporates regional score-based gradient guidance and a time travel strategy into the diffusion sampling process. Extensive experiments demonstrate that DiffEditor achieves state-of-the-art performance on various fine-grained image editing tasks, including object moving, resizing, content dragging, appearance replacing, and object pasting. The method is efficient and can be integrated into existing diffusion models without task-specific training.DiffEditor is a novel method designed to enhance the accuracy and flexibility of diffusion-based image editing. It addresses two key issues: (1) lack of editing accuracy and unexpected artifacts in complex scenarios, and (2) insufficient flexibility in harmonizing editing operations. The method introduces image prompts to fine-grained image editing, which, when combined with text prompts, provide more detailed descriptions of the editing content. To improve flexibility while maintaining content consistency, DiffEditor combines stochastic differential equations (SDE) with ordinary differential equations (ODE) sampling. Additionally, it incorporates regional score-based gradient guidance and a time travel strategy into the diffusion sampling process. Extensive experiments demonstrate that DiffEditor achieves state-of-the-art performance on various fine-grained image editing tasks, including object moving, resizing, content dragging, appearance replacing, and object pasting. The method is efficient and can be integrated into existing diffusion models without task-specific training.