MaskEdit: High Fidelity Semantic Image Editing Based on Mask Guidance

Building on the success of large scale language image models, current semantic image editing methods achieve intuitive and universal editing through diffusion models. However, most methods require providing masks to highlight the editing areas or the original text prompts of the input images, making it difficult to ensure high fidelity. This paper proposes MaskEdit, a new method that combines automatic mask generation with mask guided inversion. MaskEdit automatically generates a mask that emphasizes the editing area by comparing the diffusion model predictions under positive and negative guidance scales, allowing editing to be performed with only the original image and editing commands. Additionally, MaskEdit utilizes the automatically generated mask to guide the inversion process of the input image, adding noise within the masked area while retaining the parts outside of it, thereby improving editing fidelity. Experiments on the ImageNet dataset demonstrate that MaskEdit outperforms existing baseline methods in both editing performance and image fidelity.